Verified migration vs DIY
Pasting your repo into ChatGPT feels free.
Then you meet the anys.
A raw LLM will migrate your file and quietly reach for any and @ts-ignore every time it hits something hard, because that's the fastest way to make the errors disappear. On our public benchmark across 5 real OSS repos, Refactyl shipped 3–30× fewer anys than raw GPT-4.1-mini, with every file gated by the real compiler. See the numbers →
No credit card · Docker-sandboxed · Deleted on handoff · How we handle your code →
The honest read
“I'll just do it myself with Claude” is the most common migration plan in the world, and it works, until the repo is bigger than the context window, until file #40 contradicts the types you set in file #3, until you've spent a weekend hand-reconciling tsc errors the model swore it fixed.
The hidden cost of DIY-LLM isn't the subscription (you already pay it). It's your hours playing compiler and the any-soup you inherit forever. Refactyl does the transform and the verification: real compiler, repair loop, never-any, honest flag on anything it can't safely convert. You review a clean diff instead of debugging a hopeful one.
3–30×
less `any`/KLOC than raw GPT-4.1-mini
Across all 5 benchmark repos
0
files shipped with `@ts-ignore` or untyped `any`
Policy enforced by the engine, not the prompt
5/5
repos pass `tsc --noEmit` after Refactyl
ts-migrate fails to run at all on modern Node
Benchmark: multer, winston, debug, ws, serve-static. Pinned tags, real tsc --noEmit, reproducible. Run it yourself →
Side-by-side comparison
| Feature | Refactyl | ChatGPT / Claude (by hand) |
|---|---|---|
| any density (5-repo public benchmark) | Baseline: 1–8 any/KLOC | 3–30× more any/KLOC |
| Compiler-gated, file by file | Yes | No, you check |
| Handles repos bigger than context window | Yes | No, you chunk it manually |
| Consistency across the whole repo | Yes | Degrades as the repo grows |
| Flags what it can't safely convert | Yes | No, silently guesses |
| Total control and visibility | No | Yes |
| Nothing leaves your machine | No | Yes |
| Price | $0–$299/mo, or project quote | "Free" + your weekend |
Where each approach wins, honestly
Where Refactyl wins
- Benchmark-proven 3–30× less `any`/KLOC
- Whole-repo consistency: file #40 knows about file #3
- Compiler gate on every file, not just the ones you check
- Handles repos larger than any context window
- Honest fallbacks that flag what it can't convert, never silently guessing
- No manual chunking, re-prompting, or error reconciliation
Where DIY wins, honestly
- Total control: you decide every type, every tradeoff
- Nothing leaves your machine
- Genuinely free for a one-file tweak
- You learn the codebase deeply by doing it yourself
- No new tool to adopt if you already pay for Claude or ChatGPT
DIY is the right call for one file or a tiny package. When the repo is real, the deadline is real, and you don't want to own the any-soup forever, that's the job Refactyl was built for.
Who Refactyl is for
Anyone who priced their migration at “a weekend” and is now three weekends in.
If the migration is one file or a tiny package, paste it into Claude and move on. If it's a real repo with inter-file types, a real deadline, and a production build that needs to be clean, Refactyl does the transform and the verification so you don't have to.
Try it now
Paste Express. Get Fastify. Right here.
The same engine that migrates whole repos, scoped to a single file. Free, in your browser, no signup.
// public demo model · never paste secrets, API keys, or .env values
// fastify output appears here
// run transform to see the conversion
// free preview · 2 transforms per hour
FAQ
Should I use an LLM myself or Refactyl?
Honest answers to the real questions before you decide.
See the difference in your own repo
See the benchmark, then run yours free.
One free migration per month, up to 100 files. No credit card. The same compiler-gated, never-any engine that produced the benchmark numbers, now on your actual codebase.
No credit card · Docker-sandboxed · Deleted on handoff · Never used for training